Researched and written with 294,105 tokens (in 10 · out 19,937 · cache-read 51,511 · cache-create 222,647) across 5 Claude calls

Coding agents

• billion-context: Context-compression plugin that claims 5x fewer tokens, 100K windows, and month-long sessions with a 95-97% prefix-cache hit rate and ≤2% compression cost; directly relevant to caching and context tricks.
• Microsoft cross-platform sandboxing library: Open-source sandbox for Windows, macOS and Linux built on processcontainer, bubblewrap and seatbelt, a reusable isolation layer for running untrusted agent code.
• Cockroach Labs: bugs as patients, agents as a medical team: Five-month experiment that used a role-split agent team to add Db2 migration support to MOLT in about two days; commenters compare it to Brooks' surgical team and suggest a post-merge acceptance-test agent.
• Lee Robinson: how always-on proactive agents work: Stanford CS146S lecture covering model changes, the harness, and context engineering for always-on agents.
• t3code: Reported to let Claude orchestrate OpenAI or open-weight models from a plain conversation, with no custom subagent config.
• Claude Code on Android with per-message model routing: Local phone setup that routes each message to a different model.
• New controls in Copilot for JetBrains: Adds control over default models and the MCP server, plus easier diagnostics and chat navigation.

Skills and CLAUDE.md practice

• andrej-karpathy-skills: Single CLAUDE.md distilled from Karpathy's observations on LLM coding pitfalls, chiefly models making wrong assumptions and running with them.
• Matt Pocock on /retro: Proposes having /retro recommend turning repeated tasks into a CLI plus skill, so agents delegate to scripts they own, which means fewer tokens and fewer hallucinations.
• Matt Pocock: grill, wayfinder, or one-shot: Don't /grill-with-docs when the diff is tiny, since one-shotting is fine there, and never start with /wayfinder.
• Matt Pocock: skills still matter: Argues models are getting better at using skills too, and good skills state what you care about, then stay out of the way.
• Do you clean up your CLAUDE.md?: Reddit thread on pruning versus endlessly appending to CLAUDE.md.

CLI/TUI

• Diffle: Keyboard-driven browser diff viewer with LSP support that turns line, block or file comments into an agent prompt or a pending GitHub review; commenters say they now review PRs inside an agent harness with a review skill.
• Claude Code tab-focus on completion (Windows Terminal): Hook that switches you to the tab when Claude Code finishes, so you stop missing completions.

Agentic SaaS and self-hosting

• AgentDock: One inbox for answering Claude Code and Codex prompts from any terminal.
• API Claws by Buda: Single API for building a self-improving agent harness.
• Talorys: Personal AI agent that runs in your own Cloudflare account on the free tier; commenters dispute calling it self-hosted and one warns of Workers AI neuron billing despite the free tier.

LLM mechanics

• Retrofitting language models to operate over bytes: Nature paper on converting subword-token LLMs to byte-level operation, which would avoid the fine-grained information loss of tokenization.